Understanding Surface Water Dynamics in Post-Mining Area Through Multi-Source Remote Sensing and Spatial Regression Analysis

Highlights

What are the main findings?

• An increase in water content of vegetation and minor changes in soil moisture were noted between 2015 and 2024.
• The extent of water in the Wartowice flotation reservoir increased significantly over the last 10 years.

What is the implication of the main finding?

• Past copper mining was not the main driving factor of the observed changes in surface water according to local and global regression models.
• Remote sensing enabled detection of surface water changes over a post-mining area.

Abstract

Despite successful land reclamation efforts, post-mining areas are still prone to secondary effects of mineral extraction. These effects include surface deformations, damage to infrastructure and buildings, and periodic or permanent changes to surface water resources. This study focused on analyzing a former copper mine in southwest Poland in terms of surface water changes, which may be caused by the restoration of groundwater conditions in the region after mine closure. The main objective of the study was to detect areas with statistically significant changes in surface water between 2015 and 2024, as well as to identify the main factors influencing the observed changes. The methodology integrated open remote sensing datasets from Landsat and Sentinel-1 missions for deriving spectral indices—Modified Normalized Difference Water Index (MNDWI) and Normalized Difference Moisture Index (NDMI), as well as Surface Soil Moisture index (SSM); spatial statistics methods, including Emerging Hot Spot analysis; and regression models—Random Forest Regression (RFR) and Geographically Weighted Regression (GWR). The results obtained indicated a general increase in vegetation water content, a reduction in the extent of surface water, and minor soil moisture changes during the analyzed period. The Emerging Hot Spot analysis revealed a number of new hot spots, indicating regions with statistically significant increases in surface water content in the study area. Out of the investigated regression models, global regression (RFR) outperformed local (GWR) models, with R2 ranging between 74.7% and 87.3% for the studied dependent variables. The most important factors in terms of influence were the distance from groundwater wells, surface topography, vegetation conditions and distance from active mining areas, while surface geology conditions and permeability had the least importance in the regression models. Overall, this study offers a comprehensive framework for integrating multi-source data to support the analysis of environmental changes in post-mining regions.

1. Introduction

1.1. Surface Water Changes in Mining and Post-Mining Areas

The extraction of mineral resources leads to significant changes in the environment, affecting surface water resources in particular. Mining activities can alter the natural hydrogeological conditions of the affected areas, alter the porosity and permeability of sediments, and induce changes in topography, resulting in spatio-temporal changes to the extent and quality of surface water [1,2]. As documented in the literature, mining areas are often intensively dewatered to allow mining operations, leading to land surface subsidence. If mining workings are separated from the aquifer by a thick layer of impermeable rocks, waterlogging can occur, which leads to the formation of temporary and permanent wetlands. However, intensive water drainage commonly lowers the groundwater table, creating an extensive depression cone. This can cause the level of lakes or rivers to drop, as well as damage the alteration of rivers [2,3,4]. Areas with a lower groundwater table are also at risk of reduced water supply to the soil and root zones of plants, which can lead to lower soil moisture and leaf water content, and consequently, the degradation or deterioration of plant habitats [5]. Furthermore, active mining can contaminate surface waters with heavily polluted runoff from technological and extraction processes [6,7,8].

Spatio-temporal changes in surface water are also observed after the mining activities have been completed, often many years after the extraction has ceased. These changes in post-mining areas are attributed to the hydrogeological rebound and recovery of water conditions, which can lead to ground surface uplift (of a much smaller magnitude than the subsidence from rock mass drainage) and local flooding [9,10]. It should also be noted that the established land reclamation strategy and subsequent land use of a post-mining area can significantly alter changes in surface water extent and quality. The Przyjaźń Narodów—Szyb Babina Mine in Poland and the Central German Mining District are prime examples of former open-pit mines that have been transformed into water reservoirs, creating unique anthropogenic lakes and thereby increasing local surface water resources [11,12].

1.2. Remote Sensing Data and Spatial Statistics Methods in Monitoring Surface Water Changes

As indicated in the previous section, mining and post-mining areas represent vulnerable ecosystems due to intense interference with their geology and hydrogeology. In these environments, changes in the groundwater table can directly impact surface water resources. Conventional monitoring methods, which rely on point-based field measurements, are time-consuming, expensive, and often impractical due to limited access to certain areas. Consequently, satellite remote sensing has become an increasingly popular tool for monitoring mining and post-mining areas. It enables the observation of key water-related parameters, such as soil moisture, water extent, and vegetation water content.

A literature review reveals that passive remote sensing imagery is a primary source of information in number of studies concerning surface water in active and former mining areas. Imagery from the Visible and Near-Infrared (VNIR) and Short-Wave Infrared (SWIR) spectra has been used to identify wetlands in a reclaimed hard coal mining area [13] and to monitor surface water changes resulting from groundwater rebound in a former zinc-lead ore mine [14]. Remote sensing imagery has also been used to perform spatial modeling and assess soil moisture content in post-mining areas [15,16], monitor aquatic reclamation processes in open-pit mines [17], and separate mining-related reservoirs from natural water bodies [18,19]. Furthermore, optical satellite imagery has been used to identify spatio-temporal changes in rivers and water bodies near active mining operations [20,21] and to monitor drought effects in the Olkusz-Pomorzany post-mining area [22]. Vast datasets of Landsat images processed by the Google Earth Engine platform have been used to precisely monitor water bodies in mining areas. This monitoring can be enhanced by existing and newly developed spectral indices, such as the Improved Normalized Difference Water Index (INDWI) proposed by [23]. A comprehensive hydrological description of a post-mining area, integrating topographic, spectral, and hydrogeological indices with the Analytic Hierarchy Process (AHP), was proposed by [24]. The authors focused on detecting water bodies and rivers, determining water flow direction, assessing vegetation condition, and identifying areas of erosion, water accumulation and increased soil moisture.

In addition to their use in measuring ground surface displacements in active mining areas through Synthetic Aperture Radar Interferometry (InSAR), satellite SAR images are also essential for monitoring post-mining surface movements at millimeter or centimeter magnitudes. These techniques complement conventional survey methods and enable large-scale analysis [25,26]. InSAR-based studies of post-mining areas have demonstrated a clear relationship between observed ground uplift and groundwater table rebound in mines located in Poland [10,26], Germany [27], Belgium [28], Ukraine [29], and China [30]. The integration of InSAR results with data from other satellite missions, such as GRACE, has enabled researchers to demonstrate a relationship between ground displacement during mining activities and changes in groundwater storage capabilities. This approach allows for inverse modeling of groundwater table changes using surface displacement measurements [31]. Beyond displacement monitoring, active remote sensing has also been successfully applied to detect water bodies in active mining areas [32] and to assess soil moisture conditions [33].

The literature review indicates that spatial statistics can be utilized to describe surface water changes in mining and post-mining areas, with examples found in [24,34,35]. However, their application in this specific context remains relatively unexplored. In contrast, spatial statistics combined with remote sensing datasets are more frequently used to analyze various environmental components in mining areas, such as vegetation [36], topography [37], and soils [38].

1.3. Research Motivation and Objectives

The following research gaps were identified through the literature review presented in the previous sections:

• A limited number of studies have integrated passive and active remote sensing data for a comprehensive analysis of surface water changes, particularly in mining and post-mining areas in Europe;
• Long-term surface water changes in former or active mining areas are rarely examined using remote sensing data;
• There is a lack of research on the application of advanced spatial statistics methods (e.g., Hot Spot analysis, spatial regression, cell statistics) to study surface water dynamics.
• The main objective of this study, derived from the identified research gaps, was to identify statistically significant changes in surface water in a post-mining area over a 10-year period and to determine the factors responsible for these changes. The specific objectives of this study were as follows: (1) description of surface water changes, including the characteristics of changes in water concentration in vegetation and soils; (2) detection of areas of statistically significant changes in surface water between 2015 and 2024; and (3) identification of factors that significantly influence surface water changes using both global and local regression models.
• The study was conducted in the former copper mining area “Konrad”, located in southwest Poland. No studies that combine remote sensing data with spatial statistics and regression models have been conducted in this region to date. The “Konrad” mining area features a complex geological structure, with numerous tectonic faults significantly influencing water flow regulation in the region following mine closure, which affects changes in surface water. To describe these changes in the studied 2014–2025 period, a time series of Landsat-8/9 and Sentinel-1 SAR imagery was utilized to calculate a various surface water and moisture related indices. Identification of areas with statistically significant surface water changes was conducted using the Emerging Hot Spot analysis. To identify the factors significantly influencing the occurrence of hot or cold spots in the study area, Random Forest Regression and Geographically Weighted Regression models were applied.

2. Study Area

2.1. Location

The subject of this study was the former underground copper mining site, the Konrad Mining Plant (Zakłady Górnicze Konrad), located in the Old Copper Basin (OCB) in southwest Poland, specifically in the Warta Bolesławiecka commune (Bolesławiec County) of the Lower Silesian Voivodeship (Figure 1).

The analyzed mine is situated in a warm transitional temperate climate zone, with average temperatures of around 0 °C in winter and 18 °C in summer. Precipitation ranges from approximately 80 mm in the colder months to 225 mm in the warmer months (based on the average for the period of 1991–2020) [39]. Agricultural areas currently dominate the land cover, with small patches of tall vegetation. The area is also characterized by dispersed built-up areas, with the only larger settlements being Warta Bolesławiecka, Iwiny, and Raciborowice Dolne.

2.2. Geological and Hydrogeological Conditions

The OCB deposit is situated within the North Sudeten Basin, where the Lubichów and Konrad mines exploit the deposit found within the Grodziecka syncline. The syncline axis runs from the South–East to the North–West and dips towards the North–West. The syncline is intersected by numerous faults parallel to the basin axis, which disturb its stratigraphic structure. The strata-geological profile begins with Permian red mudstone formations overlying a crystalline foundation [40]. The copper deposit is located in the Lower and Middle Zechstein formations, approximately 18 m thick [41], with the deposit thickness not exceeding 2 m [42]. Due to the presence of faults and the North–West dip of the deposit, it has a block structure and occurs at depths ranging from 50 to 1000 m. The deposit is overlain by Upper Zechstein formations (clays, sandstones, and dolomites), followed by Triassic and Cretaceous rocks, and Quaternary sediments (Figure 2) [43].

Three water-bearing strata can be distinguished within the Grodziec syncline: Quaternary, Permian, and Mesozoic. The first stratum consists of water-bearing formations made up of sands and gravels, reaching thicknesses of up to 50 m. It is isolated from the Mesozoic stratum by tills (Figure 3).

The Mesozoic stratum is an aquifer where the Upper Cretaceous and Muschelkalk can be distinguished. It is characterized by low water saturation and hydrological interconnectivity. Along it is the Buntsandstein, which is the most abundant reservoir. However, it has low porosity and hydraulic conductivity, which limits its lack of impact on the mine’s water hazard. The Permian stratum consists of Upper Zechstein, Middle Zechstein, and Rotliegend sandstones. The Upper Zechstein is characterized by low porosity and water saturation, which is locally supplied by water from the Quaternary strata. The Middle Zechstein posed the most significant water hazard for the Konrad mine. Composed of limestones with relatively high porosity, the area is saturated with groundwater and hydrologically connected to the Quaternary strata. The Rotliegend sandstones had no impact on the water influx into the mine due to their low hydraulic conductivity [41,49].

2.3. History of Mining in the Old Copper Basin and Land Reclamation Activities

Germany initiated the exploration of the Grodziec syncline deposit by 1936. However, work related to deepening the production shafts did not start until 1938. Mine construction was interrupted by World War II, during which the completed K-II shaft and the constructed K-I, L-I, and L-II shafts were flooded. The Konrad mine was formally established in 1950. The first copper ore was extracted in 1952. In 1956, the neighboring Lubichów mine was opened, making the deposit accessible through the L-I to L-IV shafts. The two mines operated independently until 1960, when they were merged into a single entity known as the Konrad mine. Due to the presence of numerous small tectonic faults, mining at the Konrad mine was conducted in separate blocks and several levels, ranging from 50 m to a maximum of 830 m [50].

An Ore Enrichment Plant producing copper concentrate operated at the mine. Tailings were deposited in three reservoirs: Iwiny I, Iwiny II, and Wartowice (indicated on Figure 1). The construction of the Iwiny I reservoir began in 1952. After the failure of the head dam in 1967, the Iwiny II reservoir was built [49], covering an area of 12 hectares with a capacity of approximately 120,000 m³. After the reconstruction of the Iwiny I dam, it remained operational until 1971. Its area covered 129 hectares and accumulated 15,800,000 m³ of tailings. The reclamation of the reservoir toward reforestation began in 1978 [51]. However, due to the physico-chemical properties of the accumulated waste, the reservoir is still only partially covered by vegetation. The Wartowice reservoir, which operated from 1971 to 1989, covered an area of 232 hectares and stored 19,300,000 m³ of tailings. After the mine closed in 1989, there was an unsuccessful attempt at reclamation [52,53].

The Konrad mine was closed in 1989. The water drainage continued until 2001, primarily because of to the extraction of drinking water from the mine’s aquifer [54]. Following the predictions, the groundwater levels were expected to stabilize after a maximum of 8 years. However, in 2001, it was found that the flooding process occurred one-third faster than predicted [49]. Furthermore, in 2002, drinking water extraction was suspended after the water intake was contaminated with mine water, which further accelerated the restoration of the groundwater table. The Lubichów mine continued anhydrite extraction beyond 1989 until 2015.

Ground subsidence caused by the extraction at the Konrad mine reached a maximum of 1.4 m [42]. After the mine closure, between 2001 and 2007, a slight uplift of the ground surface, reaching up to 84 mm, was observed. Restoration of groundwater table was indicated as the cause for the uplift [10,55]. The uplift caused by groundwater restoration in the Lubichów mine area reached 70 mm from 2014 to 2018.

3. Materials and Methods

The methodology used in this study, illustrated in Figure 4, comprised five stages: (1) data acquisition for the study area, (2) preprocessing of satellite imagery, (3) application of selected spatial statistics, (4) development of dependent and independent variables, and (5) construction of global and local machine learning regression models. Each of these stages is discussed in detail in the following sections.

3.1. Data Acquisition

The development of models explaining surface water changes within the Konrad copper mine involved preparing several independent variables that describe the topography, mining activities, geological and hydrogeological conditions, as well as the meteorological data for the study area. The dependent variables were developed using open satellite imagery from the Sentinel-1 and Landsat-8/9 missions, along with the Emerging Hot Spot method.

Table 1. Summary of obtained datasets.

Data Type	Name	Data Characteristics	Data Format	Source
Satellite imagery	Landsat-8/9 satellite imagery	Satellite imagery acquired between 2015 and 2024 over the study area (between April and October), displaying surface reflectance (Level-2 product)	GeoTIFF	[56]
	Sentinel-1 SAR imagery	Satellite SAR imagery acquired between 2015 and 2024 over the study area, representing radar backscatter values (Ground Range Detected, GRD) and complex phase values (Single Look Complex, SLC)	GeoTIFF
Topographic data	Surface water system	Vector data with locations of water streams and lakes in the study area	Shapefile	[57]
	Absolute elevation	3D points obtained with Aerial Laser Scanning in 2020, with spatial resolution of 5 m	ASCII	[58]
	Built-up areas	10 m resolution raster data with locations of built-up areas within the study area (data validity: 2018)	GeoTIFF	[56,59]
Mining data	Tailing dams	Vector data representing locations of tailing dams over the study area, obtained through map digitization	Shapefile	[53]
	Former copper mining	Vector data representing the extent of mining workings related to copper ore extraction. Data obtained through map digitization	Shapefile	[55]
	Active mining	Vector data representing the extent of currently active mining areas, associated with anhydrite, stone, limestone, marl, sand and gravel.	Shapefile	[60]
Meteorological data	Precipitation	Raster data (0.5° resolution) representing selected meteorological data (including precipitation levels) in the study area, acquired for years 2015–2024	GeoTIFF	[56,61]
Geological and hydrogeological data	First aquifer	Hydrogeological map of Poland (scale 1:50,000) from year 2018	JPEG	[47,48]
	Geology	Detailed geological map of Poland (scale 1:50,000) from 1995 and 2005	JPEG	[44,45]
	Faults	Generalized geological map of the border region between Poland, Germany and the Czech Republic (excluding Quaternary deposits) from 2011 (scale 1:1,000,000)	PDF	[46]
	Groundwater extraction wells	Geo-environmental map of Poland from 2015 (scale 1:50,000)	JPEG	[62,63]

summarizes all the datasets obtained in the study, which were used for the development of the variables described in Section 3.4.

3.2. Remote Sensing Data Processing

3.2.1. Passive Imagery

The study utilized multispectral data acquired between 2015 and 2024 by the OLI and OLI-2 sensors onboard the Landsat-8 and -9 satellites. Images representing atmospherically corrected surface reflectance values (Level-2 product, Collection 2, Tier 1) were acquired from the Google Earth Engine platform. To exclude months with significant cloud cover and a high likelihood of snow cover, the collection was filtered to contain only images acquired between May and September. The Landsat collection, consisting of 297 images, was preprocessed using cloud and cloud shadow masking through the pixel quality attributes mask (the pixel quality attributes, stored in the QA_PIXEL raster and generated from the CFMASK algorithm, were applied), and the reflectance values were rescaled to a 0–1 range by applying appropriate scale and offset factors. A composite image was generated for each analyzed year, representing the average reflectance values of each spectral band between May and September, resulting in a total of 10 image composites.

Time series of two spectral indices were determined using the multispectral imagery dataset:

(a)

Modified Normalized Difference Water Index (MNDWI)—introduced by [64] as a modification of the original Normalized Difference Water Index (NDWI) described in [65]. MNDWI is a normalized index that utilizes reflectance acquired in the green (GREEN) and shortwave infrared (SWIR1) spectral bands, according to Formula (1):

$$\mathrm{M}\mathrm{N}\mathrm{D}\mathrm{W}\mathrm{I}=\left(\mathrm{G}\mathrm{R}\mathrm{E}\mathrm{E}\mathrm{N}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1\right)/(\mathrm{G}\mathrm{R}\mathrm{E}\mathrm{E}\mathrm{N}+\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1)$$

(1)

The index enables accurate detection of open surface water by reducing of signals from built-up areas, vegetation, and bare soil, in contrast to the NDWI. Its values range from −1 to 1, with positive values indicating the presence of surface water.

(b)

Normalized Difference Moisture Index (NDMI)—an index that combines reflectance information acquired in the near-infrared (NIR) and shortwave infrared (SWIR1) spectral ranges, as developed by [66]:

$$\mathrm{N}\mathrm{D}\mathrm{M}\mathrm{I}=\left(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1\right)/(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{S}\mathrm{W}\mathrm{I}\mathrm{R}1)$$

(2)

The primary purpose of the NDMI is to identify changes in vegetation water content. Higher values indicate a high concentration of water in vegetation.

3.2.2. Active Imagery

To identify changes in surface water content using active satellite from 2015 to 2024, a collection of Sentinel-1 Synthetic Aperture Radar (SAR) images was obtained from the Google Earth Engine platform. The acquired imagery included the Sentinel-1 Ground Range Detected (GRD) products from the ascending orbit, which were preprocessed to obtain a calibrated, ortho-rectified product through thermal noise removal, radiometric calibration, and terrain correction. A similar processing schema as for the multispectral imagery was applied: only images acquired between May and September were selected, and annual composites were generated to obtain the average radar backscatter coefficient (in the VV polarization) for each consecutive year. The σ⁰ backscatter coefficient images were resampled to 30 m resolution to match the Landsat-8/9 resolution and reduce speckle noise.

Sentinel-1 backscatter was utilized to estimate surface soil moisture dynamics using the TU Wien Change Detection Model [67]. The model exploits temporal variations in radar backscatter for each pixel, assuming that changes in backscatter are primarily due to soil moisture variations. Radar backscatter (σ⁰) is therefore scaled by minimum (${\mathsf{\sigma }}_{dry}^0$) and maximum (${\mathsf{\sigma }}_{wet}^0$) reference backscatter values, corresponding to surface backscatter under dry and wet conditions, respectively, yielding relative soil moisture index at the time of acquisition:

$$\mathrm{S}\mathrm{S}\mathrm{M}(\mathrm{t})=\left({\mathsf{\sigma }}^0\left(\mathrm{t}\right)-{\mathsf{\sigma }}_{dry}^0\right)/({\mathsf{\sigma }}_{wet}^0-{\mathsf{\sigma }}_{dry}^0)$$

(3)

An analysis of ground surface displacements was conducted to estimate surface deformations in the study area, and to assess their impact on changes in surface water conditions. A Sentinel-1 SAR dataset in the Single Look Complex (SLC) format was accessed through the Alaska Satellite Facility platform. This dataset contains 438 SLC images from the ascending acquisition orbit, spanning the period between 2015 and 2024. Surface displacements were estimated using a modified SBAS approach applied in the MintPy software package v.1.6.1. [68]. A total of 1746 differential interferograms were generated, filtered, and multilooked to reduce decorrelation effects before being spatially unwrapped using the SNAPHU algorithm [69].

The Line-of-Sight (LOS) displacement time series obtained through SBAS analysis served as an independent variable for modeling of surface water changes. Total cumulative LOS displacement (ascending orbit) was selected as the independent variable to effectively capture the total magnitude of ground deformation over the analyzed period.

3.3. Spatial Statistics Applications

3.3.1. Cell Statistics

A statistical analysis of the obtained indices (NDMI, MNDWI, and SSM) was conducted using local map algebra functions in ArcGIS Pro software (version 3.2) for the analyzed area during the 2015–2024 period. Applying the cell statistics tools facilitated the identification of the mean and range of changes for each index, along with the years in which the lowest and highest values were recorded. These analyses were conducted for each raster cell in the analyzed Konrad mine area.

3.3.2. Emerging Hot Spot Analysis

To identify trends in the calculated time series of spectral indices and relative soil moisture, the Emerging Hot Spot (EHS) Analysis, was applied using ArcGIS Pro software. The main principle of this algorithm is to determine the Getis-Ord G_i* statistic through Hot Spot Analysis, which identifies clusters of low (cold spots) and high values (hot spots) of a given spatial variable, relative to its immediate neighborhood. In the analyzed case, the statistic was determined for each bin of the space-time cube, developed on based on the input datasets. The datasets comprised time series of spectral indices and relative soil moisture, retrieved from satellite remote sensing data [70].

The theoretical principles of Hot Spot analysis, which constitutes the first stage of the Emerging Hot Spot algorithm, were described in detail in [71,72]. The spatial statistic is calculated according to Equation (4):

$${\mathrm{G}}_{\mathrm{i}}^{\mathrm{*}}=({\sum }_{\mathrm{j}=1}^{\mathrm{n}}{\mathrm{w}}_{\mathrm{i},\mathrm{j}}\times {\mathrm{x}}_{\mathrm{j}}-\overline{\mathrm{x}}\times {\sum }_{\mathrm{j}=1}^{\mathrm{n}}{\mathrm{w}}_{\mathrm{i},\mathrm{j}})/(\mathrm{S}\times \sqrt{{\displaystyle \frac{\mathrm{n}\times {\sum }_{\mathrm{j}=1}^{\mathrm{n}}{{\mathrm{w}}_{\mathrm{i},\mathrm{j}}}^2-{({\sum }_{\mathrm{j}=1}^{\mathrm{n}}{\mathrm{w}}_{\mathrm{i},\mathrm{j}})}^2}{\mathrm{n}-1}}})$$

(4)

where:

${\mathrm{x}}_{\mathrm{j}}$—attribute value for spatial feature j,

${\mathrm{w}}_{\mathrm{i},\mathrm{j}}$—spatial weight between spatial features i and j,

n—total number of spatial features,

$\overline{\mathrm{x}}$—mean value of spatial feature j, based on (5):

$$\overline{\mathrm{x}}={\sum }_{\mathrm{j}=1}^{\mathrm{n}}{\mathrm{x}}_{\mathrm{j}}/ \mathrm{n}$$

(5)

S—standard deviation, calculated using Formula (6):

$$\mathrm{S}=\sqrt{{\displaystyle \frac{{\sum }_{\mathrm{j}=1}^{\mathrm{n}}{{\mathrm{x}}_{\mathrm{j}}}^2}{\mathrm{n}}}-{\overline{\mathrm{x}}}^2}$$

(6)

Based on the calculated Getis-Ord ${\mathrm{G}}_{\mathrm{i}}^{\mathrm{*}}$ statistics for each bin and the results of the Mann–Kendall test, the Emerging Hot Spot algorithm classifies the studied area into 17 categories, which are explained in detail in [70,73]. These include areas with no identifiable trends, as well as those with a presence of new, consecutive, intensifying, or persistent cold or hot spots [70]. In the analyzed studies, the time-step interval was 1 year, and spatial weights (describing spatial relationships between features) were defined as the contiguity edges corners.

3.4. Independent and Dependent Variables Development

Through the Emerging Hot Spot analysis, three dependent variables were prepared and examined, representing statistically significant changes in surface water dynamics over the study area. These include changes in surface water extent (SurfaceWater), total vegetation water content (LeafWaterContent), and surface soil moisture (SoilMoisture).

A total of 16 independent variables were created based on the datasets described in Table 1, utilizing a number of spatial analysis techniques and raster functions, detailed in Table 2. The independent variables dataset can be split into 4 main categories, including factors that may impact surface water changes according to [2,74]:

• Topographic variables: terrain elevation (DEM2020) and slope (Slope), total Line-of-Sight displacement (DisplacementTotal), distance from rivers (DistRivers), mean NDVI value (NDVI), and distance from built-up areas (DistBuildings);
• Geological-hydrogeological variables: maximum depth of the first aquifer (MaxGroundwaterTable), the geological structure of the study area (Geology), soil permeability (Permeability), the distance from groundwater wells (DistWaterIntake), and the distance from tectonic faults (DistFault);
• Mining variables: the distance from areas of active mining (DistCurrentMining), the distance from sites associated with former copper ore mining (DistMiningCopper), and the distance from tailings dams (DistFlotation);
• Meteorological variables: average precipitation (Precipitation) and Land Surface Temperature (LST) during the studied period of 2015–2024.

All processed variables were prepared as vector point data with identical spatial extent, indicated in red in Figure 1, and spatial reference system (WGS-84 UTM Zone 33N). Points were distributed on a regular grid with 30 m spacing, corresponding to the spatial resolution of Landsat-8/9.

Table 2. Summary of obtained spatial dependent and independent variables.

Variable Type	Variable Name *	Dataset Used	Development Method	Value Range [unit]
Dependent	SurfaceWater	Landsat-8/9 imagery	Described in detail in Section 3.3 and Section 3.4	−4.20–4.05 [-]
	LeafWaterContent	Landsat-8/9 imagery	Described in detail in Section 3.3 and Section 3.4	−3.89–4.20 [-]
	SoilMoisture	Sentinel-1 imagery	Described in detail in Section 3.4	−3.94–3.94 [-]
Independent	DisplacementTotal	Sentinel-1 imagery	Described in detail in Section 3.4	−235–66 [mm]
	DistRivers	Water network data	Euclidean distance from rivers obtained through spatial analysis	0.0–2089.0 [m]
	DEM2020	ALS-based Digital Elevation Model	Digital Elevation Model (DEM) in the form of GRID model, obtained through mosaicking and resampling the XYZ point data	191.15–293.75 [m a.s.l.]
	Slope	Digital Elevation Model	Slope calculated based on the DEM2020 variable	0.0–30.3 [°]
	NDVI	Landsat-8/9 imagery	Mean NDVI [75] value during 2015–2024 estimated for Landsat imagery dataset described in Section 3.2.1.	−0.05–0.45 [-]
	DistBuildings	Built-up areas dataset	Euclidean distance from built-up areas (vectorized from raster data beforehand)	0.0–1381.3 [m]
	Precipitation	Precipitation	Conversion of raster dataset to vector format	49.38–50.05 [mm/m2]
	LST	Landsat-8/9 imagery	Land Surface Temperature obtained from cloud-free Landsat surface reflectance dataset, converted to point format	292.4–304.3 [K]
	MaxGroundwaterTable	First aquifer— hydrogeological map	Hydrogeological map digitization and conversion to raster format. Height of the first aquifer obtained through subtracting aquifer depth from surface elevation	145.29–269.67 [m a.s.l.]
	Permeability	Geological maps	Geological map digitization and conversion to raster format. Permeability determined based on types of quaternary deposits, applying distribution provided in [76]	[1]—excellent; [2]—good; [3]—average; [4]—weak; [5]—semi-pervious; [6]—impervious; [0]—anthropogenic areas
	Geology	Geological maps	Geological maps digitization and raster conversion	[0]—basalts; [1]—clays and deluvial sands; [2]—tills; [3]—loesses and loess-like clays; [4]—silts; [5]—muds; [6]—unknown; [7]—sands and gravels; [8]—sandstones; [9]—limestones and marls
	DistFault	Faults-geological maps	Digitization of the generalized geological map of the border region between Poland, Germany and the Czech Republic; Euclidean distance from faults	0.0–1770.0 [m]
	DistWaterIntake	Groundwater wells locations	Digitization of the geo-environmental map of Poland; Euclidean distance from wells	84.8–9436.3 [m]
	DistCurrentMining	Active mining	Euclidean distance from active mining areas	0.0–4883.5 [m]
	DistFlotation	Tailings dams	Euclidean distance from tailings dams locations	0.0–3545.6 [m]
	DistMiningCopper	Former copper ore mining	Euclidean distance from copper ore fields	0.0–2305.3 [m]

* Labels: blue—topographic variable, red—mining variable, brown—geological-hydrogeological variable, orange—meteorological variable.

Table 2. Summary of obtained spatial dependent and independent variables.

Variable Type	Variable Name *	Dataset Used	Development Method	Value Range [unit]
Dependent	SurfaceWater	Landsat-8/9 imagery	Described in detail in Section 3.3 and Section 3.4	−4.20–4.05 [-]
	LeafWaterContent	Landsat-8/9 imagery	Described in detail in Section 3.3 and Section 3.4	−3.89–4.20 [-]
	SoilMoisture	Sentinel-1 imagery	Described in detail in Section 3.4	−3.94–3.94 [-]
Independent	DisplacementTotal	Sentinel-1 imagery	Described in detail in Section 3.4	−235–66 [mm]
	DistRivers	Water network data	Euclidean distance from rivers obtained through spatial analysis	0.0–2089.0 [m]
	DEM2020	ALS-based Digital Elevation Model	Digital Elevation Model (DEM) in the form of GRID model, obtained through mosaicking and resampling the XYZ point data	191.15–293.75 [m a.s.l.]
	Slope	Digital Elevation Model	Slope calculated based on the DEM2020 variable	0.0–30.3 [°]
	NDVI	Landsat-8/9 imagery	Mean NDVI [75] value during 2015–2024 estimated for Landsat imagery dataset described in Section 3.2.1.	−0.05–0.45 [-]
	DistBuildings	Built-up areas dataset	Euclidean distance from built-up areas (vectorized from raster data beforehand)	0.0–1381.3 [m]
	Precipitation	Precipitation	Conversion of raster dataset to vector format	49.38–50.05 [mm/m2]
	LST	Landsat-8/9 imagery	Land Surface Temperature obtained from cloud-free Landsat surface reflectance dataset, converted to point format	292.4–304.3 [K]
	MaxGroundwaterTable	First aquifer— hydrogeological map	Hydrogeological map digitization and conversion to raster format. Height of the first aquifer obtained through subtracting aquifer depth from surface elevation	145.29–269.67 [m a.s.l.]
	Permeability	Geological maps	Geological map digitization and conversion to raster format. Permeability determined based on types of quaternary deposits, applying distribution provided in [76]	[1]—excellent; [2]—good; [3]—average; [4]—weak; [5]—semi-pervious; [6]—impervious; [0]—anthropogenic areas
	Geology	Geological maps	Geological maps digitization and raster conversion	[0]—basalts; [1]—clays and deluvial sands; [2]—tills; [3]—loesses and loess-like clays; [4]—silts; [5]—muds; [6]—unknown; [7]—sands and gravels; [8]—sandstones; [9]—limestones and marls
	DistFault	Faults-geological maps	Digitization of the generalized geological map of the border region between Poland, Germany and the Czech Republic; Euclidean distance from faults	0.0–1770.0 [m]
	DistWaterIntake	Groundwater wells locations	Digitization of the geo-environmental map of Poland; Euclidean distance from wells	84.8–9436.3 [m]
	DistCurrentMining	Active mining	Euclidean distance from active mining areas	0.0–4883.5 [m]
	DistFlotation	Tailings dams	Euclidean distance from tailings dams locations	0.0–3545.6 [m]
	DistMiningCopper	Former copper ore mining	Euclidean distance from copper ore fields	0.0–2305.3 [m]

* Labels: blue—topographic variable, red—mining variable, brown—geological-hydrogeological variable, orange—meteorological variable.

3.5. Development of Global and Local Regression Models

The relationships between the dependent and independent variables were analyzed using a global regression modeling approach. The Random Forest Regression (RFR) method, first introduced by [77], is based on an ensemble of decision trees. Each decision tree performs regression analysis independently, with the final outcome of the algorithm being the mean of the results from all decision trees. Each decision tree is trained using a random subset of training points and independent variables that define the splits within the trees. These features make RFR more resistant to overfitting, less sensitive to changes in the input dataset, and eliminate the need for tree pruning [78,79,80]. A Random Forest Regression algorithm implemented in the Scikit-Learn Python v. 1.7.0. library was selected for this study. Through the grid search, the optimal hyperparameters were estimated: 200 estimators (decision trees), maximum tree depth of 20, and minimum samples per split of 5. The model was trained on 80% of data samples, and its evaluation was carried out on the remaining 20%.

Local regression modeling was carried out using the Geographically Weighted Regression (GWR) method. Unlike global models, the non-parametric GWR model allows the relationships between the variables to vary across space, therefore accounting for spatial heterogeneity in these relationships. The equation for the GWR model is given by (7):

$${\mathrm{y}}_{\mathrm{i}}\left(\mathrm{u}\right)={\mathsf{\beta }}_{0\mathrm{i}}\left(\mathrm{u}\right)+{\mathsf{\beta }}_{1\mathrm{i}}\left(\mathrm{u}\right)\times {\mathrm{a}}_1+{\mathsf{\beta }}_{2\mathrm{i}}\left(\mathrm{u}\right)\times {\mathrm{a}}_2+\cdots +{\mathsf{\beta }}_{\mathrm{n}\mathrm{i}}\left(\mathrm{u}\right)\times {\mathrm{a}}_{\mathrm{n}}+{\mathsf{\epsilon }}_{\mathrm{i}}$$

(7)

where

${\mathrm{y}}_{\mathrm{i}}\left(\mathrm{u}\right)$—value of the dependent variable at location u,

${\mathsf{\beta }}_{0\mathrm{i}}\left(\mathrm{u}\right)$—intercept of the regression equation at location u,

${\mathsf{\beta }}_{1\mathrm{i}}\left(\mathrm{u}\right), {{\mathsf{\beta }}_{2\mathrm{i}}\left(\mathrm{u}\right),\dots ,\mathsf{\beta }}_{\mathrm{n}\mathrm{i}}\left(\mathrm{u}\right)$—coefficients describing the relationships between the dependent variable and factors ${\mathrm{a}}_1, {\mathrm{a}}_2,\dots ,$ ${\mathrm{a}}_{\mathrm{n}}$ at location u,

${\mathsf{\epsilon }}_{\mathrm{i}}$—random error at location u [81].

GWR develops local models based on a spatial neighborhood matrix, using a weighting function [82]. More information on the GWR model can also be found in [83,84].

The local GWR models in this study were developed using ArcGIS Pro (version 3.2) software. Due to the discrete distribution of the Geology and Permeability variables, the models did not include these factors. A local multicollinearity analysis, based on the Variance Inflation Factor (VIF) and the distribution of local standard deviation values of individual independent variables, excluded the DistCurrentMining, DistMiningCopper, DistWaterIntake, MaxGroundwaterTable, NDVI, Precipitation, and Slope factors (the observed VIF values were higher than 5.4 and the local standard deviation values for certain regions were close to 0.0). Given the normal distribution of the dependent variables, a Gaussian model was adopted and a Gaussian kernel weighting was applied [85]. The study area was divided into irregular neighborhoods using the Golden Search approach.

Accuracy of both the RFR and GWR models was assessed. The R², Root Mean Squared Error (RMSE), and Mean Absolute Error (MAE) metrics were used for the RFR approach, and the GWR models were evaluated using R² and Akaike Information Criterion corrected (AICc) [86,87,88]. Additionally, histograms of regression residuals were prepared to verify their distribution.

4. Results

4.1. Analysis of Spatio-Temporal Changes in Surface Water Conditions

4.1.1. Cell Statistics

The cell statistics evaluated in the study included the mean index values from 2015 to 2024, the range of changes, and the years of lowest and highest index values.

The statistical analysis of the MNDWI time series revealed that minimum and maximum index values were observed during early and late years of the studied period, depending on the location within the study area (Figure 5a,b). In the southern part of the study area, the lowest index values were observed between 2020 and 2024, except for the central part of the Wartowice tailings dam, where the lowest values were observed in 2015. In the northern part of the study area, low MNDWI values were noted during 2017–2020. In the west, a majority of analyzed cells had minimal values in 2024. After 2020, the highest MNDWI values were recorded in the north–west, central, and south–east parts of the study area, as well as within the Wartowice tailings dam.

Figure 5c,d display the mean MNDWI values alongside the range of index change, respectively. The analysis, spanning 2015–2024, indicated that the highest mean values (approximately 0.6) occurred in the Wartowice reservoir, as well as in the central part of the study area and near Warta Bolesławiecka. The lowest mean values (between −0.6 and −0.4) were found in the northwest, east, and northeast, corresponding to the high density of vegetation cover. The MNDWI range of change was between 0.01 and 1.26, with the most significant changes (higher than 0.8) observed in areas with the highest mean values and in the former mining areas of the Lubichów mine.

Figure 6 shows the cell statistics of the NDMI time series. Figure 6a,b indicate that the lowest vegetation water content was observed in 2015 and 2021. In contrast, the maximum index values varied both temporally and spatially across the study area. The western part of the study area recorded its highest NDMI values in 2015 and 2016, while the southern and eastern parts reached index peaks in 2021. In the north, a significant number of cells reached peak values in 2023 and 2024.

Mean NDMI values are presented in Figure 6c. The highest mean values, which exceeded 0.45, were observed in the eastern and northern parts of the study area. The lowest values, below −0.1, were recorded in the northern part of the Wartowice tailings dam and in the vicinity of the Iwiny I and Iwiny II reservoirs. In the 2015–2024 period, the NDMI was subject to changes ranging from 0.01 to 0.90 (Figure 6d). The most significant changes (higher than 0.65) were observed in the southern part of the Wartowice reservoir and within agricultural areas. The northern part of the Wartowice reservoir, as well as the urban areas and the eastern part of the study area, experienced minimal changes (less than 0.15).

Figure 7 displays the cell statistics for the soil moisture index time series, with inland waters and built-up areas masked during the Sentinel-1 data processing. Based on Figure 7a,b both the minimum and maximum values of the SSM index were observed primarily until 2019, except for small areas in the northwestern, central, and southern parts of the study area. The peak values of the index in these areas were observed since 2022.

Analysis of the pixel values in Figure 7c reveals that the highest mean SSM index values (greater than 0.6) occurred in the eastern section of the former copper ore mining area. The lowest mean values were found in the northern part of the Wartowice tailings dam, within the Iwiny I and Iwiny II reservoirs, as well as in the north-western part of the study area. Over the study period, the index values varied within a range of 0.03 to 0.87 (Figure 7d). The most significant changes were observed east of Raciborowice Dolne, along with several small areas across the study site. Areas adjacent to the Wartowice reservoir were also characterized by substantial changes.

4.1.2. Emerging Hot Spot Analysis

The results of the Emerging Hot Spot analysis for the MNDWI time series, shown on Figure 8, indicate the presence of Persistent Cold Spots within the Konrad mine area, which correspond to consistently low index values over the study period (2015–2024). These areas, characterized by tall vegetation, were primarily identified in the eastern, northern, and north-eastern parts of the study area. The analysis also revealed several Persistent Hot Spots (zones with consistently high MNDWI values) located in the southern part of the Wartowice reservoir, within the Iwiny I reservoir, and in the western and southern parts of the study area. Several New Hot Spots were observed in the northwestern part, near the former Lubichów plant, where MNDWI values increased between 2015 and 2024, indicating an increase in surface water content. Small clusters of Intensifying Hot Spots were also identified within the Wartowice reservoir and in the southern part of the study area, indicating a systematic upward trend in the MNDWI observations.

Figure 9 presents the Emerging Hot Spot analysis for the NDMI time series, indicating a general increase in vegetation water content across the study area. Numerous Persistent Hot Spots, representing consistently high NDMI values, were identified throughout the eastern, northern, and northeastern parts of the Konrad mine. New Hot Spots were observed in the neighborhood of the Wartowice reservoir and in the northern part of the area of interest (AOI), indicating a recent increase in vegetation water content during 2015–2024. It should be noted that the sites of Iwiny I, Iwiny II, and Wartowice tailings dams were classified as Persistent Cold Spots, where there was no significant increase in vegetation water content, and the NDMI values remained low throughout the study period.

Results of the Emerging Hot Spot analysis for the SSM index are shown in Figure 10. The map confirms areas of significant and persistent soil moisture (Persistent Hot Spots), and an increasing moisture content trend (Consecutive Hot Spots) in the eastern, northern, and north-eastern areas of the Konrad mine. Higher SSM values were also observed near the Wartowice reservoir and in the southern part of the AOI. Finally, the analysis detected numerous New and Intensifying Cold Spots, primarily in the western and central parts of the post-mining area and in the northern part of the Wartowice reservoir. This indicates a decrease in the soil moisture index over the study period.

4.2. Impact of Independent Variables on Surface Water Changes—Global-Scale Approach

Table 3. Accuracy metrics for the considered global RFR models.

Dependent Variable	Metric
	R2 [%]	RMSE	MAE
SurfaceWater	78.6	0.679	0.490
LeafWaterContent	87.3	0.442	0.301
SoilMoisture	74.7	0.645	0.437

presents the accuracy metrics for the Random Forest Regression (RFR) models developed in this study. The metrics indicate that the LeafWaterContent model achieved the best fit, accounting for 87.3% of the statistically significant variance in vegetation water content across the study area between 2015 and 2024, with an RMSE of 0.442. The other two models, the SurfaceWater and SoilMoisture models, exhibit slightly lower R² values of 78.6% and 74.7%, respectively. Their corresponding RMSE were 0.679 and 0.645, indicating lower accuracy than the best-performing model. It is important to note that the regression residuals for all considered RFR models follow a normal distribution, as confirmed by the histograms attached in Appendix A.

Figure 11 summarizes the impact of the independent variables on statistically significant changes in surface water extent, vegetation water content, and soil moisture, based on the results of RFR modeling. Figure 11a reveals that, for global regression, the SurfaceWater variable was most impacted by following explanatory variables: DistWaterIntake (importance = 12.2%), DistCurrentMining (importance = 10.6%), and DEM2020 (importance = 10.0%), indicating that the vicinity of groundwater wells and active mining areas, as well as surface topography, have the highest impact on surface water changes. On the other hand, geological conditions and permeability had the least impact on surface water changes in the Konrad mine area.

The impact of predictor variables on the LeafWaterContent variable is shown in Figure 11b. The analysis confirms the significant impact (22.9%) of the NDVI on the observed change in vegetation water content within the study area during 2015–2024. This reflects the improving condition of vegetation in the post-mining area, as well as the inherent correlation between the NDVI and NDMI indices derived from the same Landsat-8/9 imagery. Furthermore, distance from groundwater wells (DistWaterIntake) and from active mining operations (DistCurrentMining) had a significant influence in the RFR modeling, with variable importances of over 13.0% and 9.0%, respectively. Similarly to the SurfaceWater model, geology was the least significant of the independent variables.

Figure 11c indicates that, among the analyzed variables, the SoilMoisture model was most influenced by the DistCurrentMining (11.6%), DistWaterIntake (11.3%) and DEM2020 (10.6%) variables. The least important were the Geology, Permeability and Precipitation variables, with importance values not exceeding 0.5%.

Although the Random Forest Regression models do not identify former copper mining as the primary driver of environmental changes in surface water, its influence remains significant. The importance of the distance from flotation dams (DistFlotation) was over 7.0% for the LeafWaterContent model, over 9.0% for the SurfaceWater model, and 9.4% for the SoilMoisture model. The impact from former copper mining sites was slightly lower, reaching 6.4%, 5.7% and 4.0% for the respective models.

4.3. Impact of Independent Variables on Surface Water Changes—Local-Scale Approach

Accuracy metrics for the Geographically Weighted Regression (GWR) models are presented in

Table 4. Accuracy metrics for the local GWR models.

Dependent Variable	Metric
	R2 [%]	AICc	Number of Neighbors
SurfaceWater	75.9	96361.08	31
LeafWaterContent	83.0	67616.91	31
SoilMoisture	68.4	96295.63	31

. Through the interpretation of the accuracy metrics, the local modeling approach did not improve the explanatory power compared to the global models. The GWR model for the SurfaceWater, LeafWaterContent and SoilMoisture variables yielded R² values that were lower than the corresponding RFR models by 2.7%, 4.3% and 6.3%, respectively. The lower-than-expected performance of GWR modeling is likely due to the exclusion of variables of a discrete character and exhibiting strong local correlations. Furthermore, the local regression models are more complex, as indicated by high Akaike Information Criterion (AICc) values. The regression residuals do not always follow a normal distribution; for instance, the residuals of the LeafWaterContent model are slightly left skewed, as indicated in the Appendix A.

Table 5. Summary of β coefficients assigned to independent variables in GWR models.

Dependent Variable	Independent Variable	Diagnostic			Area of the Strongest Impact (|β| ≥ 1.5)
		Min	Max	StDev
SurfaceWater	DEM2020	−4.00	2.40	0.25	Wartowice tailings dam with its immediate surroundings, areas located east of Warta Bolesławiecka town, area of the Iwiny I flotation reservoir
	DistBuildings	−0.54	0.28	0.02	not applicable
	DistFault	−0.30	0.59	0.02	not applicable
	DistFlotation	−0.26	0.16	0.02	not applicable
	DistRivers	−0.21	0.26	0.02	not applicable
	LST	−2.39	2.54	0.48	western part of Lubków town, eastern part of Warta Bolesławiecka, central, south-eastern and north-eastern part of study area
	DisplacementTotal	−0.31	0.17	0.01	n/a
LeafWaterContent	DEM2020	−2.08	1.82	0.20	the eastern part of Warta Bolesławiecka, Wartowice tailings dam, areas north-east of the Iwiny I reservoir
	DistBuildings	−0.21	0.19	0.01	not applicable
	DistFault	−0.20	0.24	0.01	not applicable
	DistFlotation	−0.18	0.22	0.01	not applicable
	DistRivers	−0.20	0.18	0.01	not applicable
	LST	−2.44	1.60	0.36	southern and north-eastern parts of the study area
	DisplacementTotal	−0.42	0.16	0.01	not applicable
SoilMoisture	DEM2020	−3.10	1.57	0.22	Wartowice and Iwiny I tailings dams
	DistBuildings	−0.33	0.66	0.02	not applicable
	DistFault	−0.73	0.35	0.02	not applicable
	DistFlotation	−0.30	0.51	0.02	not applicable
	DistRivers	−0.46	0.30	0.02	not applicable
	LST	−2.45	2.16	0.43	eastern part of the study area
	DisplacementTotal	−0.22	0.11	0.01	not applicable

summarizes the beta (β) coefficients for the independent variables in each GWR model, highlighting areas with the strongest impacts, which correspond to absolute coefficient values exceeding 1.5. The results indicate that the DEM2020 and LST had the most significant local impact on surface water changes. The DistFault variable had a significant influence only in the SoilMoisture model, with coefficients ranging from −0.73 to 0.35. On the other hand, the lowest influence was observed for the DistFlotation, DistRivers and DisplacementTotal variables.

An analysis of the spatial distribution of the β coefficients revealed that only the DEM2020 and LST exhibited significant local influence (less than −1.5 or greater than 1.5) in specific areas of the study. These locations include the eastern part of Warta Bolesławiecka, areas of the Wartowice and Iwiny I tailings dams, and the eastern, southeastern, and southwestern parts of the AOI, as indicated in Appendix B. For the remaining of independent variables, no significant local impacts were identified.

5. Discussion

This section elaborates on several aspects of the research methodology and obtained the results that warrant further discussion. One aspect is the use of Landsat-8/9 imagery to derive the MNDWI and NDMI indices, despite the operational availability of Sentinel-2 data with a higher spatial resolution of 10 m since 2015. This choice was driven by the use of the Google Earth Engine (GEE) platform for data processing, where Sentinel-2 surface reflectance products have been available since March 2017. To ensure a radiometrically consistent time series for the analyzed period, Landsat imagery was used instead. Furthermore, although the GEE platform has built-in atmospheric correction algorithms, these were not applied, because they do not align with the official procedures used by the Copernicus ground segment. It should also be emphasized that objects present in post-mining areas (e.g., tailing dams) may affect surface reflectance, generated using the USGS Land Surface Reflectance Code (LaSRC) algorithm, and thus the values of the MNDWI and NDMI indices. However, the results obtained in this study indicate that the changes in spectral indices values observed in the research area were significantly higher than the reflectance uncertainties documented in publication [89].

The MNDWI and NDMI spectral indices were selected to describe surface water changes, based on literature review conducted in Section 1.2. These indices are the most common for detecting water bodies and assessing vegetation water content in mining regions [22,23]. The Surface Soil Moisture (SSM) index was calculated using SAR data from the ascending orbit, correspondingly to the ascending orbit-based displacement measurements from InSAR. Future research should also incorporate data from the descending orbit for a comparative analysis of the relative SSM values. However, it should also be noted that the SSM estimation methodology applied in this paper is highly sensitive to extreme values of radar backscatter in areas covered by vegetation, buildings and inland water. To mitigate the effects, built-up areas and water bodies were masked during the preprocessing of SAR data. A temporal smoothing was also applied through the estimation of annual SSM time series, further minimizing the potential impact of anomalous values. The authors acknowledge that despite these efforts some residual errors may still be present and therefore affect the final SSM evaluation, particularly in areas of heterogeneous vegetation cover and on tailing surfaces. While localized errors may influence pixel-level estimates, this analysis focused on relative temporal changes and aggregated spatial patterns, which are less sensitive to these uncertainties. Nonetheless, future work should further address this limitation by exploring alternative approaches for soil moisture analysis, incorporating ancillary datasets and validating remote sensing results with in situ soil moisture measurements for improved accuracy.

In this study, annual composites presenting average reflectance values for each band from May to September were used to determine spectral indices. The main purpose of this procedure was to identify global trends in surface water changes in the specified post-mining area and to exclude months with unfavorable weather conditions, such as cloud cover and snowfall. Furthermore, it was essential in the context of the EHS analysis, which requires a constant time-step interval between data of time series, as confirmed by research [73,90]. However, it would be valuable in the future to analyze seasonal changes in surface water resources using satellite data with high radiometric quality and low cloud cover. Additionally, the EHS method could be applied for each year separately.

A diversity of spatial resolutions could pose a problem for regression-based analysis of variable impact, particularly concerning Precipitation independent variable. In consequence, this variable was excluded from the GWR modeling. Furthermore, the accuracy of the geological, hydrogeological, and mining-related variables was likely influenced by the georeferencing and digitization process of the analog source materials.

A total of 16 independent variables were used in this study to explain statistically significant changes in surface water by examining water indices and their relationships through regression modeling. The accuracy of global and local models could be further enhanced by incorporating additional explanatory variables describing the former copper mining in greater detail, e.g., the depth of the underground mining fields. Moreover, environmental and meteorological factors, as well as human activity, could have influenced changes in surface water resources. However, these factors could not be considered in this analysis due to the lack of relevant geospatial data (e.g., air temperature and distance from hydrotechnical constructions). Another improvement would be the inclusion of the Line-of-Sight (LOS) displacement decomposition from both ascending and descending orbits in the InSAR processing chain. This would enable information on total vertical ground surface changes for the regression models.

It should also be mentioned that prior to selecting the Random Forest Regression (RFR) for modeling in the global regression approach, several other machine learning algorithms were tested, including Extreme Gradient Boosting (XGBoost), Support Vector Regression (SVR), and linear regression with regularization (Ridge, Lasso and ElasticNet Regression). Since the accuracy of the RFR model was significantly better than that of the other methods, only the RFR results are presented in this paper (a detailed accuracy assessment of the tested global regression models is provided in Table S1). Despite their high accuracy and good fit to the data, the RFR residuals exhibited spatial clustering, as confirmed by the global Moran’s I statistics.

The results presented in this study demonstrate that global RFR models better explain statistically significant changes in surface water between 2015 and 2024 than local GWR models. A potential explanation for these findings may be the exclusion of several independent variables from the modeling process, justified by their local multicollinearity or a discrete form. In addition, local relationships occurring in the study area may be minor or non-linear, making them difficult to identify using GWR models. The distribution of the SurfaceWater and LeafWaterContent dependent variables warrants further discussion, as these variables did not follow a perfectly normal distribution, which could not be achieved even with a Box–Cox transform [91]. This may have impacted accuracy of GWR models, since GWR employs a Gaussian regression model type. A potential approach for future studies would be to transform the dependent variables into a binary format (e.g., 0—no change, 1—significant change), and apply a corresponding regression model type (e.g., Logistic Regression) to analyze the local relationships between surface water and environmental variables.

The applicability and relevance of this research would be substantially increased by validating the results with in situ data, such as groundwater table measurements from local piezometers. Knowledge of groundwater table level fluctuations during the studied 2015–2024 period would greatly aid in understanding the observed surface water changes. Unfortunately, despite the presence of numerous piezometers in the study area, the data is not publicly available. However, to prove the relevance of the research, the average annual precipitation values for the years 2015–2024 (analyzed only from May to September for each year), recorded by the Twardocice station, were compared with the average values of the MNDWI, NDMI and SSM indices in three selected regions (Figure S1). The graph confirms the lack of significant correlation between precipitation amounts and remote sensing indices, highlighting the need for studies that consider additional geological, hydrogeological, mining and topographical variables.

6. Conclusions

This study analyzed the spatio-temporal changes in surface water, vegetation water content, and soil moisture in a post-mining area in southwest Poland from 2015 to 2024. A time series of the MNDWI, NDMI, and SSM indices were developed using both multispectral (Landsat-8/9) and Synthetic Aperture Radar (Sentinel-1) satellite imagery to characterize these changes and create a set of dependent variables for regression analysis. Selected cell statistics tools were applied to offer a comprehensive overview of the changes in water conditions over the study area. Furthermore, the Emerging Hot Spot analysis method was used to identify locations of statistically significant trends during the 10-year period. Finally, by applying global Random Forest Regression (RFR) and local Geographically Weighted Regression (GWR) modeling, a set of factors contributing to the formation of hot and cold spots were identified in the former copper ore mine area.

The obtained results indicated an overall increase in vegetation water content, a decrease in surface water extent (with an exception of the Wartowice tailings dam), and minor changes in soil moisture across the study area. The Emerging Hot Spot analysis for the MNDWI time series revealed numerous intensifying and new hot spots within the Wartowice tailings reservoir and the former Lubichów mining plant. These regions also contained new, intensifying, and consecutive hot spots in the NDMI and SSM time series. In contrast, the eastern part of the study area featured extensive areas of persistent hot spots for the NDMI and SSM indices, indicating high water content in vegetation and soils throughout the 10-year period studied in this paper. However, this region, characterized by tall vegetation, also exhibited persistent cold spots for the MNDWI.

The global models, which ignore the spatial relationships between variables, proved more accurate than the local models (which consider the spatial characteristics) by approximately 2.7–6.3%, depending on the considered variable. The most accurate model developed was the model explaining statistically significant changes in vegetation water content, with an R² score of 87.3%. The GWR model for the soil moisture index was the least accurate, with an R² equal to 68.4%. The analysis of feature importances from the RFR models and the β coefficients from the GWR models revealed that former copper mining was not the primary driver of the observed surface water changes in the study area. The most influential factors, identified from the studied explanatory variables, were the distance to groundwater wells, surface elevation, distance to current mining activities, and overall vegetation conditions. The least significant variables were the location of geological deposits and the permeability of geological strata. In the local modeling, land surface temperature was also identified as a significant explanatory factor.

This research demonstrates the significant potential of data fusion through connecting multi-source, multi-resolution data from both passive and active satellite remote sensing, geospatial data and statistical data for long-term analysis and monitoring of surface water changes in post-mining areas. Nevertheless, several areas warrant further investigation, including the application of high-resolution remote sensing data, incorporating a larger number of independent variables that better explain the former mining activities, and attempting to elucidate surface water changes using other spectral and radar-based remote sensing indices as dependent variables. Such studies could reveal a potentially greater impact of former mining activities on surface water dynamics, allowing for these changes to be monitored and described with greater precision.